Electricity Price Forecasting with Principal Component-Guided Sparse Regression

This study investigates forecasting day-ahead electricity prices in markets influenced by weather-dependent energy sources. We utilize the principal component-guided sparse regression (pcLasso) technique, developed by Tay et al., to manage multicollinearity among weather and grid variables effectively. Our focus is on the Japanese Electricity Power Exchange (JEPX), which integrates significant solar capacity, demonstrating how various weather variables impact electricity prices. Our results show that pcLasso surpasses state-of-the-art models like Lasso and Ridge regression in stability and consistency of parameter estimation. Importantly, pcLasso enhances forecast accuracy as more variables are incorporated, providing insights into the seasonality and inter-annual variations of weather impacts. These findings underscore pcLasso's ability to generate robust, interpretable models for complex forecasting scenarios. A key finding when applying pcLasso to JEPX price is the significant improvement in accuracy using day-ahead reserve margin forecasts over weekly forecasts, highlighting the importance of timely disclosure of grid information.